DAPD: Dual-Anchored Policy Distillation
The paper argues that policy distillation can teach a student model to rely on information it will not have at inference time.
The authors call this failure mode “privilege illusion” and trace it to asymmetry between a privileged teacher and the deployed student. DAPD adds two anchoring mechanisms meant to align behavior under matched-information paths and reduce overreliance on privileged reference guidance. In the reported experiments, it beats OPSD on Qwen3-4B by 2.00 points on average, with larger gains listed at 4B and 32B scales. HF Daily Papers' note
The authors call this failure mode “privilege illusion” and trace it to asymmetry between a privileged teacher and the deployed student. DAPD adds two anchoring mechanisms meant to align behavior under matched-information paths and reduce overreliance on privileged reference guidance. In the reported experiments, it beats OPSD on Qwen3-4B by 2.00 points on average, with larger gains listed at 4B and 32B scales. HF Daily Papers' note
score 5